CycleGAN, Image-to-Image Translation¶
In this notebook, we're going to define and train a CycleGAN to read in an image from a set $X$ and transform it so that it looks as if it belongs in set $Y$. Specifically, we'll look at a set of images of Yosemite national park taken either during the summer of winter. The seasons are our two domains!
The objective will be to train generators that learn to transform an image from domain $X$ into an image that looks like it came from domain $Y$ (and vice versa).
Some examples of image data in both sets are pictured below.
Unpaired Training Data¶
These images do not come with labels, but CycleGANs give us a way to learn the mapping between one image domain and another using an unsupervised approach. A CycleGAN is designed for image-to-image translation and it learns from unpaired training data. This means that in order to train a generator to translate images from domain $X$ to domain $Y$, we do not have to have exact correspondences between individual images in those domains. For example, in the paper that introduced CycleGANs, the authors are able to translate between images of horses and zebras, even though there are no images of a zebra in exactly the same position as a horse or with exactly the same background, etc. Thus, CycleGANs enable learning a mapping from one domain $X$ to another domain $Y$ without having to find perfectly-matched, training pairs!
CycleGAN and Notebook Structure¶
A CycleGAN is made of two types of networks: discriminators, and generators. In this example, the discriminators are responsible for classifying images as real or fake (for both $X$ and $Y$ kinds of images). The generators are responsible for generating convincing, fake images for both kinds of images.
This notebook will detail the steps one should take to define and train such a CycleGAN.
- Load in the image data using PyTorch's DataLoader class to efficiently read in images from a specified directory.
- Then, define the CycleGAN architecture according to provided specifications, the discriminator and the generator models.
- Complete the training cycle by calculating the adversarial and cycle consistency losses for the generator and discriminator network and completing a number of training epochs.
- Finally, evaluate your model by looking at the loss over time and looking at sample, generated images.
Load and Visualize the Data¶
We'll first load in and visualize the training data, importing the necessary libraries to do so.
!unzip summer2winter_yosemite.zip # can comment out after executing once
# loading in and transforming data
import os
import torch
from torch.utils.data import DataLoader
import torchvision
import torchvision.datasets as datasets
import torchvision.transforms as transforms
# visualizing data
import matplotlib.pyplot as plt
import numpy as np
import warnings
%matplotlib inline
DataLoaders¶
The get_data_loader function returns training and test DataLoaders that can load data efficiently and in specified batches. The function has the following parameters:
image_type:summerorwinter, the names of the directories where the X and Y images are storedimage_dir: name of the main image directory, which holds all training and test imagesimage_size: resized, square image dimension (all images will be resized to this dim)batch_size: number of images in one batch of data
The test data is strictly for feeding to our generators, later on, so we can visualize some generated samples on fixed, test data.
You can see that this function is also responsible for making sure our images are of the right, square size (128x128x3) and converted into Tensor image types.
It's suggested that you use the default values of these parameters.
Note: If you are trying this code on a different set of data, you may get better results with larger image_size and batch_size parameters. If you change the batch_size, make sure that you create complete batches in the training loop otherwise you may get an error when trying to save sample data.
def get_data_loader(image_type, image_dir='summer2winter_yosemite',
image_size=128, batch_size=16, num_workers=0):
"""Returns training and test data loaders for a given image type, either 'summer' or 'winter'.
These images will be resized to 128x128x3, by default, converted into Tensors, and normalized.
"""
# resize and normalize the images
transform = transforms.Compose([transforms.Resize(image_size), # resize to 128x128
transforms.ToTensor()])
# get training and test directories
image_path = './' + image_dir
train_path = os.path.join(image_path, image_type)
test_path = os.path.join(image_path, 'test_{}'.format(image_type))
# define datasets using ImageFolder
train_dataset = datasets.ImageFolder(train_path, transform)
test_dataset = datasets.ImageFolder(test_path, transform)
# create and return DataLoaders
train_loader = DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers)
test_loader = DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers)
return train_loader, test_loader
# Create train and test dataloaders for images from the two domains X and Y
# image_type = directory names for our data
dataloader_X, test_dataloader_X = get_data_loader(image_type='summer')
dataloader_Y, test_dataloader_Y = get_data_loader(image_type='winter')
Display some Training Images¶
Below we provide a function imshow that reshape some given images and converts them to NumPy images so that they can be displayed by plt. This cell should display a grid that contains a batch of image data from set $X$.
# helper imshow function
def imshow(img):
npimg = img.numpy()
plt.imshow(np.transpose(npimg, (1, 2, 0)))
# get some images from X
dataiter = iter(dataloader_X)
# the "_" is a placeholder for no labels
images, _ = dataiter.next()
# show images
fig = plt.figure(figsize=(12, 8))
imshow(torchvision.utils.make_grid(images))
Next, let's visualize a batch of images from set $Y$.
# get some images from Y
dataiter = iter(dataloader_Y)
images, _ = dataiter.next()
# show images
fig = plt.figure(figsize=(12,8))
imshow(torchvision.utils.make_grid(images))
Pre-processing: scaling from -1 to 1¶
We need to do a bit of pre-processing; we know that the output of our tanh activated generator will contain pixel values in a range from -1 to 1, and so, we need to rescale our training images to a range of -1 to 1. (Right now, they are in a range from 0-1.)
# current range
img = images[0]
print('Min: ', img.min())
print('Max: ', img.max())
# helper scale function
def scale(x, feature_range=(-1, 1)):
''' Scale takes in an image x and returns that image, scaled
with a feature_range of pixel values from -1 to 1.
This function assumes that the input x is already scaled from 0-255.'''
# scale from 0-1 to feature_range
min, max = feature_range
x = x * (max - min) + min
return x
# scaled range
scaled_img = scale(img)
print('Scaled min: ', scaled_img.min())
print('Scaled max: ', scaled_img.max())
Define the Model¶
A CycleGAN is made of two discriminator and two generator networks.
Discriminators¶
The discriminators, $D_X$ and $D_Y$, in this CycleGAN are convolutional neural networks that see an image and attempt to classify it as real or fake. In this case, real is indicated by an output close to 1 and fake as close to 0. The discriminators have the following architecture:
This network sees a 128x128x3 image, and passes it through 5 convolutional layers that downsample the image by a factor of 2. The first four convolutional layers have a BatchNorm and ReLu activation function applied to their output, and the last acts as a classification layer that outputs one value.
Convolutional Helper Function¶
To define the discriminators, we make use of this conv function, which creates a convolutional layer + an optional batch norm layer.
import torch.nn as nn
import torch.nn.functional as F
# helper conv function
def conv(in_channels, out_channels, kernel_size, stride=2, padding=1, batch_norm=True):
"""Creates a convolutional layer, with optional batch normalization.
"""
layers = []
conv_layer = nn.Conv2d(in_channels=in_channels, out_channels=out_channels,
kernel_size=kernel_size, stride=stride, padding=padding, bias=False)
layers.append(conv_layer)
if batch_norm:
layers.append(nn.BatchNorm2d(out_channels))
return nn.Sequential(*layers)
Define the Discriminator Architecture¶
Tthe __init__ function has the specified 5 layer conv net architecture. Both $D_X$ and $D_Y$ have the same architecture, so we only need to define one class, and later instantiate two discriminators.
It's recommended to use a kernel size of 4x4 and use that to determine the correct stride and padding size for each layer. This Stanford resource may also help in determining stride and padding sizes.
- Define your convolutional layers in
__init__ - Then fill in the forward behavior of the network
The forward function defines how an input image moves through the discriminator, and the most important thing is to pass it through your convolutional layers in order, with a ReLu activation function applied to all but the last layer.
To reiterate, we do not apply a sigmoid activation function to the output, here, and that is because we are planning on using a squared error loss for training. And you can read more about this loss function, later in the notebook.
class Discriminator(nn.Module):
def __init__(self, conv_dim=64):
super(Discriminator, self).__init__()
self.conv1 = conv(3, conv_dim, 4, batch_norm=False)
self.conv2 = conv(conv_dim, conv_dim * 2, 4)
self.conv3 = conv(conv_dim * 2, conv_dim * 4, 4)
self.conv4 = conv(conv_dim * 4, conv_dim * 8, 4)
self.conv5 = conv(conv_dim * 8, 1, 4, batch_norm=False)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = F.relu(self.conv3(x))
x = F.relu(self.conv4(x))
x = self.conv5(x)
return x
Generators¶
The generators, G_XtoY and G_YtoX (sometimes called F), are made of an encoder, a conv net that is responsible for turning an image into a smaller feature representation, and a decoder, a transpose_conv net that is responsible for turning that representation into an transformed image. These generators, one from XtoY and one from YtoX, have the following architecture:
This network sees a 128x128x3 image, compresses it into a feature representation as it goes through three convolutional layers and reaches a series of residual blocks. It goes through a few (typically 6 or more) of these residual blocks, then it goes through three transpose convolutional layers (sometimes called de-conv layers) which upsample the output of the resnet blocks and create a new image!
Note that most of the convolutional and transpose-convolutional layers have BatchNorm and ReLu functions applied to their outputs with the exception of the final transpose convolutional layer, which has a tanh activation function applied to the output. Also, the residual blocks are made of convolutional and batch normalization layers, which we'll go over in more detail, next.
Residual Block Class¶
To define the generators, you're expected to define a ResidualBlock class which will help you connect the encoder and decoder portions of the generators. You might be wondering, what exactly is a Resnet block? It may sound familiar from something like ResNet50 for image classification, pictured below.
ResNet blocks rely on connecting the output of one layer with the input of an earlier layer. The motivation for this structure is as follows: very deep neural networks can be difficult to train. Deeper networks are more likely to have vanishing or exploding gradients and, therefore, have trouble reaching convergence; batch normalization helps with this a bit. However, during training, we often see that deep networks respond with a kind of training degradation. Essentially, the training accuracy stops improving and gets saturated at some point during training. In the worst cases, deep models would see their training accuracy actually worsen over time!
One solution to this problem is to use Resnet blocks that allow us to learn so-called residual functions as they are applied to layer inputs. You can read more about this proposed architecture in the paper, Deep Residual Learning for Image Recognition by Kaiming He et. al, and the below image is from that paper.
Residual Functions¶
Usually, when we create a deep learning model, the model (several layers with activations applied) is responsible for learning a mapping, M, from an input x to an output y.
M(x) = y(Equation 1)
Instead of learning a direct mapping from x to y, we can instead define a residual function
F(x) = M(x) - x
This looks at the difference between a mapping applied to x and the original input, x. F(x) is, typically, two convolutional layers + normalization layer and a ReLu in between. These convolutional layers should have the same number of inputs as outputs. This mapping can then be written as the following; a function of the residual function and the input x. The addition step creates a kind of loop that connects the input x to the output, y:
M(x) = F(x) + x(Equation 2) or
y = F(x) + x(Equation 3)
Optimizing a Residual Function¶
The idea is that it is easier to optimize this residual function F(x) than it is to optimize the original mapping M(x). Consider an example; what if we want y = x?
From our first, direct mapping equation, Equation 1, we could set M(x) = x but it is easier to solve the residual equation F(x) = 0, which, when plugged in to Equation 3, yields y = x.
Defining the ResidualBlock Class¶
To define the ResidualBlock class, we'll define residual functions (a series of layers), apply them to an input x and add them to that same input. This is defined just like any other neural network, with an __init__ function and the addition step in the forward function.
In our case, we define the residual block as:
- Two convolutional layers with the same size input and output
- Batch normalization applied to the outputs of the convolutional layers
- A ReLu function on the output of the first convolutional layer
Then, in the forward function, add the input x to this residual block.
# residual block class
class ResidualBlock(nn.Module):
"""Defines a residual block.
This adds an input x to a convolutional layer (applied to x) with the same size input and output.
These blocks allow a model to learn an effective transformation from one domain to another.
"""
def __init__(self, conv_dim):
super(ResidualBlock, self).__init__()
self.conv1 = conv(conv_dim, conv_dim, kernel_size=3, stride=1, padding=1, batch_norm=True)
self.conv2 = conv(conv_dim, conv_dim, kernel_size=3, stride=1, padding=1, batch_norm=True)
def forward(self, x):
# apply a ReLu activation the outputs of the first layer
# return a summed output, x + resnet_block(x)
x = F.relu(self.conv1(x))
x = x + self.conv2(x)
return x
import torch.nn as nn
import torch.nn.functional as F
# helper block function
def blocks(conv_dim, number_of_blocks = 6):
"""Creates a residual block of size `conv_dim`
"""
layers = []
for layer in range(number_of_blocks):
layers.append(ResidualBlock(conv_dim))
return nn.Sequential(*layers)
Transpose Convolutional Helper Function¶
To define the generators, we use the above conv function, ResidualBlock class, and the below deconv helper function, which creates a transpose convolutional layer + an optional batchnorm layer.
# helper deconv function
def deconv(in_channels, out_channels, kernel_size, stride=2, padding=1, batch_norm=True):
"""Creates a transpose convolutional layer, with optional batch normalization.
"""
layers = []
# append transpose conv layer
layers.append(nn.ConvTranspose2d(in_channels, out_channels, kernel_size, stride, padding, bias=False))
# optional batch norm layer
if batch_norm:
layers.append(nn.BatchNorm2d(out_channels))
return nn.Sequential(*layers)
Define the Generator Architecture¶
- Complete the
__init__function with the specified 3 layer encoder convolutional net, a series of residual blocks (the number of which is given byn_res_blocks), and then a 3 layer decoder transpose convolutional net. - Then complete the
forwardfunction to define the forward behavior of the generators. Recall that the last layer has atanhactivation function.
Both $G_{XtoY}$ and $G_{YtoX}$ have the same architecture, so we only need to define one class, and later instantiate two generators.
class CycleGenerator(nn.Module):
def __init__(self, conv_dim=64, n_res_blocks=6):
super(CycleGenerator, self).__init__()
# 1. Define the encoder part of the generator
# initial convolutional layer given, below
self.conv1 = conv(3, conv_dim, 4)
self.conv2 = conv(conv_dim, conv_dim*2, 4)
self.conv3 = conv(conv_dim*2, conv_dim*4, 4)
# 2. Define the resnet part of the generator
# Residual blocks
self.blocks = blocks(conv_dim * 4, n_res_blocks)
# 3. Define the decoder part of the generator
# two transpose convolutional layers and a third that looks a lot like the initial conv layer
self.deconv1 = deconv(conv_dim*4, conv_dim*2, 4)
self.deconv2 = deconv(conv_dim*2, conv_dim, 4)
# no batch norm on last layer
self.deconv3 = deconv(conv_dim, 3, 4, batch_norm=False)
def forward(self, x):
"""Given an image x, returns a transformed image."""
# define feedforward behavior, applying activations as necessary
out = F.relu(self.conv1(x))
out = F.relu(self.conv2(out))
out = F.relu(self.conv3(out))
out = self.blocks(out)
out = F.relu(self.deconv1(out))
out = F.relu(self.deconv2(out))
# tanh applied to last layer
out = F.tanh(self.deconv3(out))
return out
Create the complete network¶
Using the classes defined earlier, we can define the discriminators and generators necessary to create a complete CycleGAN. The given parameters should work for training.
First, create two discriminators, one for checking if $X$ sample images are real, and one for checking if $Y$ sample images are real. Then the generators. Instantiate two of them, one for transforming a painting into a realistic photo and one for transforming a photo into into a painting.
def create_model(g_conv_dim=64, d_conv_dim=64, n_res_blocks=6):
"""Builds the generators and discriminators."""
# Instantiate generators
G_XtoY = CycleGenerator(conv_dim=g_conv_dim, n_res_blocks=n_res_blocks)
G_YtoX = CycleGenerator(conv_dim=g_conv_dim, n_res_blocks=n_res_blocks)
# Instantiate discriminators
D_X = Discriminator(conv_dim=d_conv_dim)
D_Y = Discriminator(conv_dim=d_conv_dim)
# move models to GPU, if available
if torch.cuda.is_available():
device = torch.device("cuda:0")
G_XtoY.to(device)
G_YtoX.to(device)
D_X.to(device)
D_Y.to(device)
print('Models moved to GPU.')
else:
print('Only CPU available.')
return G_XtoY, G_YtoX, D_X, D_Y
# call the function to get models
G_XtoY, G_YtoX, D_X, D_Y = create_model()
Check that you've implemented this correctly¶
The function create_model should return the two generator and two discriminator networks. After defining these discriminator and generator components, it's good practice to check our work. The easiest way to do this is to print out your model architecture and read through it to make sure the parameters are what you expected. The next cell will print out their architectures.
# helper function for printing the model architecture
def print_models(G_XtoY, G_YtoX, D_X, D_Y):
"""Prints model information for the generators and discriminators.
"""
print(" G_XtoY ")
print("-----------------------------------------------")
print(G_XtoY)
print()
print(" G_YtoX ")
print("-----------------------------------------------")
print(G_YtoX)
print()
print(" D_X ")
print("-----------------------------------------------")
print(D_X)
print()
print(" D_Y ")
print("-----------------------------------------------")
print(D_Y)
print()
# print all of the models
print_models(G_XtoY, G_YtoX, D_X, D_Y)
Discriminator and Generator Losses¶
Computing the discriminator and the generator losses are key to getting a CycleGAN to train.
Image from original paper by Jun-Yan Zhu et. al.
The CycleGAN contains two mapping functions $G: X \rightarrow Y$ and $F: Y \rightarrow X$, and associated adversarial discriminators $D_Y$ and $D_X$. (a) $D_Y$ encourages $G$ to translate $X$ into outputs indistinguishable from domain $Y$, and vice versa for $D_X$ and $F$.
To further regularize the mappings, we introduce two cycle consistency losses that capture the intuition that if we translate from one domain to the other and back again we should arrive at where we started. (b) Forward cycle-consistency loss and (c) backward cycle-consistency loss.
Least Squares GANs¶
We've seen that regular GANs treat the discriminator as a classifier with the sigmoid cross entropy loss function. However, this loss function may lead to the vanishing gradients problem during the learning process. To overcome such a problem, we'll use a least squares loss function for the discriminator. This structure is also referred to as a least squares GAN or LSGAN, and you can read the original paper on LSGANs, here. The authors show that LSGANs are able to generate higher quality images than regular GANs and that this loss type is a bit more stable during training!
Discriminator Losses¶
The discriminator losses will be mean squared errors between the output of the discriminator, given an image, and the target value, 0 or 1, depending on whether it should classify that image as fake or real. For example, for a real image, x, we can train $D_X$ by looking at how close it is to recognizing and image x as real using the mean squared error:
out_x = D_X(x)
real_err = torch.mean((out_x-1)**2)Generator Losses¶
Calculating the generator losses will look somewhat similar to calculating the discriminator loss; there will still be steps in which you generate fake images that look like they belong to the set of $X$ images but are based on real images in set $Y$, and vice versa. You'll compute the "real loss" on those generated images by looking at the output of the discriminator as it's applied to these fake images; this time, your generator aims to make the discriminator classify these fake images as real images.
Cycle Consistency Loss¶
In addition to the adversarial losses, the generator loss terms will also include the cycle consistency loss. This loss is a measure of how good a reconstructed image is, when compared to an original image.
Say you have a fake, generated image, x_hat, and a real image, y. You can get a reconstructed y_hat by applying G_XtoY(x_hat) = y_hat and then check to see if this reconstruction y_hat and the orginal image y match. For this, we recommed calculating the L1 loss, which is an absolute difference, between reconstructed and real images. You may also choose to multiply this loss by some weight value lambda_weight to convey its importance.
The total generator loss will be the sum of the generator losses and the forward and backward cycle consistency losses.
Define Loss Functions¶
To help us calculate the discriminator and gnerator losses during training, let's define some helpful loss functions. Here, we'll define three.
real_mse_lossthat looks at the output of a discriminator and returns the error based on how close that output is to being classified as real. This should be a mean squared error.fake_mse_lossthat looks at the output of a discriminator and returns the error based on how close that output is to being classified as fake. This should be a mean squared error.cycle_consistency_lossthat looks at a set of real image and a set of reconstructed/generated images, and returns the mean absolute error between them. This has alambda_weightparameter that will weight the mean absolute error in a batch.
It's recommended that you take a look at the original, CycleGAN paper to get a starting value for lambda_weight.
def real_mse_loss(D_out):
# how close is the produced output from being "real"?
real_err = torch.mean((D_out-1)**2)
return real_err
def fake_mse_loss(D_out):
# how close is the produced output from being "false"?
real_err = torch.mean((D_out)**2)
return real_err
def cycle_consistency_loss(real_im, reconstructed_im, lambda_weight):
# calculate reconstruction loss
reconstruction_loss = torch.mean(torch.abs(real_im - reconstructed_im))
# return weighted loss
return lambda_weight * reconstruction_loss
Define the Optimizers¶
Next, let's define how this model will update its weights. This, like the GANs you may have seen before, uses Adam optimizers for the discriminator and generator. It's again recommended that you take a look at the original, CycleGAN paper to get starting hyperparameter values.
import torch.optim as optim
# hyperparams for Adam optimizers
lr=0.0002
beta1=0.5
beta2=0.999
g_params = list(G_XtoY.parameters()) + list(G_YtoX.parameters()) # Get generator parameters
# Create optimizers for the generators and discriminators
g_optimizer = optim.Adam(g_params, lr, [beta1, beta2])
d_x_optimizer = optim.Adam(D_X.parameters(), lr, [beta1, beta2])
d_y_optimizer = optim.Adam(D_Y.parameters(), lr, [beta1, beta2])
Training a CycleGAN¶
When a CycleGAN trains, and sees one batch of real images from set $X$ and $Y$, it trains by performing the following steps:
Training the Discriminators
- Compute the discriminator $D_X$ loss on real images
- Generate fake images that look like domain $X$ based on real images in domain $Y$
- Compute the fake loss for $D_X$
- Compute the total loss and perform backpropagation and $D_X$ optimization
- Repeat steps 1-4 only with $D_Y$ and your domains switched!
Training the Generators
- Generate fake images that look like domain $X$ based on real images in domain $Y$
- Compute the generator loss based on how $D_X$ responds to fake $X$
- Generate reconstructed $\hat{Y}$ images based on the fake $X$ images generated in step 1
- Compute the cycle consistency loss by comparing the reconstructions with real $Y$ images
- Repeat steps 1-4 only swapping domains
- Add up all the generator and reconstruction losses and perform backpropagation + optimization
Saving Your Progress¶
A CycleGAN repeats its training process, alternating between training the discriminators and the generators, for a specified number of training iterations. You've been given code that will save some example generated images that the CycleGAN has learned to generate after a certain number of training iterations. Along with looking at the losses, these example generations should give you an idea of how well your network has trained.
Below, you may choose to keep all default parameters.
# import save code
from helpers import save_samples, checkpoint
# train the network
def training_loop(dataloader_X, dataloader_Y, test_dataloader_X, test_dataloader_Y,
n_epochs=1000):
print_every=10
# keep track of losses over time
losses = []
test_iter_X = iter(test_dataloader_X)
test_iter_Y = iter(test_dataloader_Y)
# Get some fixed data from domains X and Y for sampling. These are images that are held
# constant throughout training, that allow us to inspect the model's performance.
fixed_X = test_iter_X.next()[0]
fixed_Y = test_iter_Y.next()[0]
fixed_X = scale(fixed_X) # make sure to scale to a range -1 to 1
fixed_Y = scale(fixed_Y)
# batches per epoch
iter_X = iter(dataloader_X)
iter_Y = iter(dataloader_Y)
batches_per_epoch = min(len(iter_X), len(iter_Y))
for epoch in range(1, n_epochs+1):
# Reset iterators for each epoch
if epoch % batches_per_epoch == 0:
iter_X = iter(dataloader_X)
iter_Y = iter(dataloader_Y)
images_X, _ = iter_X.next()
images_X = scale(images_X) # make sure to scale to a range -1 to 1
images_Y, _ = iter_Y.next()
images_Y = scale(images_Y)
# move images to GPU if available (otherwise stay on CPU)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
images_X = images_X.to(device)
images_Y = images_Y.to(device)
# ============================================
# TRAIN THE DISCRIMINATORS
# ============================================
## First: D_X, real and fake loss components ##
d_x_optimizer.zero_grad()
# 1. Compute the discriminator losses on real images
out_x = D_X(images_X)
D_X_real_loss = real_mse_loss(out_x)
# 2. Generate fake images that look like domain X based on real images in domain Y
fake_X = G_YtoX(images_Y)
# 3. Compute the fake loss for D_X
out_x = D_X(images_X)
D_X_fake_loss = fake_mse_loss(out_x)
# 4. Compute the total loss and perform backprop
d_x_loss = D_X_real_loss + D_X_fake_loss
d_x_loss.backward()
d_x_optimizer.step()
## Second: D_Y, real and fake loss components ##
d_y_optimizer.zero_grad()
out_y = D_Y(images_Y)
D_Y_real_loss = real_mse_loss(out_y)
fake_Y = G_XtoY(images_X)
out_y = D_Y(images_Y)
D_Y_fake_loss = fake_mse_loss(out_y)
d_y_loss = D_Y_real_loss + D_Y_fake_loss
d_y_loss.backward()
d_y_optimizer.step()
# =========================================
# TRAIN THE GENERATORS
# =========================================
## First: generate fake X images and reconstructed Y images ##
g_optimizer.zero_grad()
# 1. Generate fake images that look like domain X based on real images in domain Y
fake_X = G_YtoX(images_Y)
# 2. Compute the generator loss based on domain X
out_x = D_X(fake_X)
g_YtoX_loss = real_mse_loss(out_x)
# 3. Create a reconstructed y
# 4. Compute the cycle consistency loss (the reconstruction loss)
reconstructed_Y = G_XtoY(fake_X)
reconstructed_y_loss = cycle_consistency_loss(images_Y, reconstructed_Y, lambda_weight=10)
## Second: generate fake Y images and reconstructed X images ##
fake_Y = G_XtoY(images_X)
out_y = D_Y(fake_Y)
g_XtoY_loss = real_mse_loss(out_y)
reconstructed_X = G_XtoY(fake_Y)
reconstructed_x_loss = cycle_consistency_loss(images_X, reconstructed_X, lambda_weight=10)
# 5. Add up all generator and reconstructed losses and perform backprop
g_total_loss = g_XtoY_loss + g_YtoX_loss + reconstructed_y_loss + reconstructed_x_loss
g_total_loss.backward()
g_optimizer.step()
# Print the log info
if epoch % print_every == 0:
# append real and fake discriminator losses and the generator loss
losses.append((d_x_loss.item(), d_y_loss.item(), g_total_loss.item()))
print('Epoch [{:5d}/{:5d}] | d_X_loss: {:6.4f} | d_Y_loss: {:6.4f} | g_total_loss: {:6.4f}'.format(
epoch, n_epochs, d_x_loss.item(), d_y_loss.item(), g_total_loss.item()))
sample_every=100
# Save the generated samples
if epoch % sample_every == 0:
G_YtoX.eval() # set generators to eval mode for sample generation
G_XtoY.eval()
save_samples(epoch, fixed_Y, fixed_X, G_YtoX, G_XtoY, batch_size=16)
G_YtoX.train()
G_XtoY.train()
# uncomment these lines, if you want to save your model
checkpoint_every=1000
# Save the model parameters
if epoch % checkpoint_every == 0:
checkpoint(epoch, G_XtoY, G_YtoX, D_X, D_Y)
return losses
n_epochs = 8000 # keep this small when testing if a model first works, then increase it to >=1000
losses = training_loop(dataloader_X, dataloader_Y, test_dataloader_X, test_dataloader_Y, n_epochs=n_epochs)
Tips on Training and Loss Patterns¶
A lot of experimentation goes into finding the best hyperparameters such that the generators and discriminators don't overpower each other. It's often a good starting point to look at existing papers to find what has worked in previous experiments, I'd recommend this DCGAN paper in addition to the original CycleGAN paper to see what worked for them. Then, you can try your own experiments based off of a good foundation.
Discriminator Losses¶
When you display the generator and discriminator losses you should see that there is always some discriminator loss; recall that we are trying to design a model that can generate good "fake" images. So, the ideal discriminator will not be able to tell the difference between real and fake images and, as such, will always have some loss. You should also see that $D_X$ and $D_Y$ are roughly at the same loss levels; if they are not, this indicates that your training is favoring one type of discriminator over the other and you may need to look at biases in your models or data.
Generator Loss¶
The generator's loss should start significantly higher than the discriminator losses because it is accounting for the loss of both generators and weighted reconstruction errors. You should see this loss decrease a lot at the start of training because initial, generated images are often far-off from being good fakes. After some time it may level off; this is normal since the generator and discriminator are both improving as they train. If you see that the loss is jumping around a lot, over time, you may want to try decreasing your learning rates or changing your cycle consistency loss to be a little more/less weighted.
fig, ax = plt.subplots(figsize=(12,8))
losses = np.array(losses)
plt.plot(losses.T[0], label='Discriminator, X', alpha=0.5)
plt.plot(losses.T[1], label='Discriminator, Y', alpha=0.5)
plt.plot(losses.T[2], label='Generators', alpha=0.5)
plt.title("Training Losses")
plt.legend()
Evaluate the Result!¶
It is useful to see how well the generated images are after an $x$ amount of epochs. This gives us a way to see whether or not the Generators are creating good fake images. For example, the image below depicts real images in the $Y$ set, and the corresponding generated images during different points in the training process. You can see that the generator starts out creating very noisy, fake images, but begins to converge to better representations as it trains (though, not perfect).
Below is a helper function for displaying generated samples based on the passed in training iteration.
import matplotlib.image as mpimg
# helper visualization code
def view_samples(iteration, sample_dir='samples_cyclegan'):
# samples are named by iteration
path_XtoY = os.path.join(sample_dir, 'sample-{:06d}-X-Y.png'.format(iteration))
path_YtoX = os.path.join(sample_dir, 'sample-{:06d}-Y-X.png'.format(iteration))
# read in those samples
try:
x2y = mpimg.imread(path_XtoY)
y2x = mpimg.imread(path_YtoX)
except:
print('Invalid number of iterations.')
fig, (ax1, ax2) = plt.subplots(figsize=(18,20), nrows=2, ncols=1, sharey=True, sharex=True)
ax1.imshow(x2y)
ax1.set_title('X to Y')
ax2.imshow(y2x)
ax2.set_title('Y to X')
# view samples at iteration 100
view_samples(100, 'samples_cyclegan')
# view samples at iteration 1000
view_samples(1000, 'samples_cyclegan')
# view samples at iteration 1000
view_samples(4000, 'samples_cyclegan')
# view samples at iteration 1000
view_samples(8000, 'samples_cyclegan')
Further Challenges and Directions¶
- One shortcoming of this model is that it produces fairly low-resolution images; this is an ongoing area of research; you can read about a higher-resolution formulation that uses a multi-scale generator model, in this paper.
- Relatedly, we may want to process these as larger (say 256x256) images at first, to take advantage of high-res data.
- It may help your model to converge faster, if you initialize the weights in your network.
- This model struggles with matching colors exactly. This is because, if $G_{YtoX}$ and $G_{XtoY}$ may change the tint of an image; the cycle consistency loss may not be affected and can still be small. You could choose to introduce a new, color-based loss term that compares $G_{YtoX}(y)$ and $y$, and $G_{XtoY}(x)$ and $x$, but then this becomes a supervised learning approach.
- This unsupervised approach also struggles with geometric changes, like changing the apparent size of individual object in an image, so it is best suited for stylistic transformations.
- For creating different kinds of models or trying out the Pix2Pix Architecture, this Github repository which implements CycleGAN and Pix2Pix in PyTorch is a great resource.
Different datasets for download¶
You can download a variety of datasets used in the Pix2Pix and CycleGAN papers, by following instructions in the associated Github repository. You'll just need to make sure that the data directories are named and organized correctly to load in that data.